A Method For Detection Mass In Mammograms Based On Multi-resolution Image Pyramids

Breast cancer is one of the most dangerous malignant diseases to the women in the world. And the breast mass is one of the main symptoms of the breast lesion. The mammogram is the most reliable clinical method to detect breast diseases. But the physicians must read lots of mammograms and in mammograms masses are usually difficult to detect as they often superimpose on dese structured background. This makes the finding of anomalous region less than 0.1% in mammograms by naked eyes is difficlty extremely. In order to solve this problem, comptuer aided diagnosis system based digital image processing has gradually become focus of the medical image.As the exploration in this study, an approach is presented in the dissertation. The approach is to detect the seed regions of the masses based on multi-resolution image pyramid and implement the location and the final description of the masses in original mammograms through the growth algorithm. The main contents are as follows:1. The multi-resolution image pyramids is the hierarchical data structure maked by images of different resolution.These images are obtained through some mathematica operations. The resolution of evey layer of pyramids declines by 2 factor. The main methods to product the image pyramids are the weighted class and decomposition class. After compared two kinds of the methods we choose the multi-resolution analysis based on wavelet theory finally.2. The neural network is the practical technology in the field of modern pattem recognition, and its application fields become widely day by day. The dissertation introduced the concept and the characteristics of the Backward Propagation (BP) neural network. Following that is the detailed detection steps to detect seed regions of the masses in low resolution images.3. After the seed-regions of masses were detected in the low resolution image, track the pixels having the same attribute pixel according to the related attributes of the pixel. We use the link rules of the overlapping type and the new discrimination rules of the right value to makes the growing process do not restricted by the area and shape of the seeds no longer. The location and the final description of the supervised regions (instead of masses region) were completed in the original images finally.4. In order to decrease the rate of false-positive, we extracted the texture features from the samples to distinguish the masses and the normal regions. We study on Angular Second Moment,Sum of Entropy,Contrast and Inverse Difference Moment based on the gray level co-occurrence and the statistics of these. Twelve statistics are chosen finally to build the neural network classifier. Use this method to process the supervised regions, and obtained the final detection results that when the rate of false-positive is 21.2% the lesions detection rate is 76.7%.